Open Collections

Manson, Matthew Alexander

University of British Columbia

Background: In Rwanda, a two-way-SMS-based mHealth intervention, WelTel, was deployed to support isolated COVID-19 patients throughout the pandemic. Patients received daily open-ended check-in-messages throughout their isolation period. To inform further health system digitalization, we sought to quantify WelTel enrollment, assess patients’ usage patterns, and explore how patient characteristics influence such behaviours. We further sought to investigate patient isolation experiences by AI-enhanced-analysis (Natural Language Processing (NLP)) of patient-clinician conversations, to improve similar programs. Methods: WelTel registration and messaging records were extracted, supplemented with Rwanda Ministry of Health data, and quantified. Patient use (≥1 conversation) was computed and compared across sociodemographic groups (sex, age, province, COVID-19-status, pandemic-wave) using logistic regression. Conversation counts and characteristics (language, messages/conversation) were quantified alongside patient communication behaviours (conversations/user, response-times) which were also compared across sociodemographic groups using non-parametric tests. To understand isolation experiences, conversations were sampled (n=2,791/12,119), English-translated (as necessary), topic-labelled, language-restored, and used to train single-topic classifiers (Traditional-ML/Transformer architectures). Best-performing models meeting a F1≥0.7 cutoff were applied to unlabeled conversations. Topic prevalence and sociodemographic differences were assessed in human-labelled, and human-and-machine- labelled corpora using logistic regression. Results: Rwanda registered 33,081 individuals in WelTel (March 2020-March 2022). Of those, 18% (n=6,021) used WelTel, with variation by sex, COVID-19-status, province, and pandemic-wave (p<0.001), but not age. 12,119 conversations were undertaken in Kinyarwanda (67%), English (25%), and regional languages. Most conversations contained <5 messages (75%). 56% of users produced one conversation (range:1-18). The median response-time was 77 minutes (IQR:22-294). Conversations/user and response-times were similar across sociodemographic groups. For conversation topic classification, traditional-ML models generally performed best and 67% of topics were suitably classified. Medical topics (e.g., symptoms:(70%), diagnostics:(37%)) were frequent. Service quality (15%), social (15%), and lifestyle (6%) topics were also discussed, alongside other rare topics. Sociodemographic factors accounted for minor differences in topics discussed. Conclusion: Rwanda’s WelTel deployment was used by a significant number of patients during the COVID-19 pandemic. NLP methods are a promising means of multilingual conversation analysis but require further optimization. Interactive texting enabled isolated patients and providers to discuss medical and non-medical issues and obtain advice that may have helped them self-manage their isolation.

Text; Dataset

2024-03-28

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/87658

Experimental Medicine

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/